Add -no-rebuild options for test-one and promote-one

[ncik-roberts]

Add these targets for running a single test without incurring the cost of a dune build:

make test-one-no-rebuild TEST=...
make promote-one-no-rebuild TEST=...

This allows tests to run ~instantly rather than waiting for upward of a minute for make test-one. You need to build the tree before you can run make test-one-no-rebuild.

There is surely a more principled way to do this (like, make test-one shouldn't take that long if no files change), but this is a quick fix that improves my workflow by a lot.

[ncik-roberts]

Some timings:

The original run:

time make test-one TEST=typing-local/local.ml
real	1m49.869s
user	7m54.249s
sys	1m34.341s

The second run:

time make test-one TEST=typing-local/local.ml
real	0m8.756s
user	0m5.930s
sys	0m2.840s

With no rebuild:

time make test-one-no-rebuild TEST=typing-local/local.ml
real	0m0.826s
user	0m0.767s
sys	0m0.061s

I walk back my original claim that test-one takes over a minute if no files changed; it's really only 8s. But I'd much rather wait 0.8s than 8s. Also sometimes I change files and want to re-run tests without caring whether I pick up my new changes or not; in this case test-one-no-rebuild does save you a full minute.

[goldfirere]

🎉 🎉 🎉

You would think my low tolerance for workflow warts would translate to a high level of interest in resolving those warts, but you'd be wrong. However, it does translate to high level of gratitude toward others who do. :)

But can I ask for a shorter name? Actually, I imagine it would be best to have something like initial-test-one for the old test-one, and then this new one can just be test-one. Or maybe even test1, but that might be overoptimized. Is the error situation if you try to run a test without building easy enough to diagnose?

[ncik-roberts]

My main counterpoint is that test-one-no-rebuild is only two characters longer than test-one: - and <TAB>.

But also, I see this new target as a convenient hack. The current test-one does the expected thing of rebuilding the compiler if you make a change, and the new target does not. I see that as a feature of test-one, not a bug. I wouldn't want to rename the new target to test-one and have people lose debugging time when they don't realize that it doesn't do a rebuild.

I also don't think the error on a null build is particularly intelligible:

$ make test-one-no-rebuild TEST=ocaml/testsuite/tests/basic/eval_order_4.ml 
(export OCAMLSRCDIR=$(pwd)/_runtest; \
 export CAML_LD_LIBRARY_PATH=$(pwd)/_runtest/lib/ocaml/stublibs; \
 if $(which gfortran > /dev/null 2>&1); then \
   export LIBRARY_PATH=$(dirname $(gfortran -print-file-name=libgfortran.a)); \
 fi; \
 cd _runtest/testsuite && make one TEST=tests/basic/eval_order_4.ml )
bash: line 5: cd: _runtest/testsuite: No such file or directory
make: *** [ocaml/Makefile.common-jst:325: test-one-no-rebuild] Error 1

Or rather, I see the thing that makes this error intelligible the fact that the command line clearly says "no-rebuild", and I have a slight opinion against removing that.

[goldfirere]

OK.

More out of curiosity than anything else: if I change the test case and use -no-rebuild, will it see my changes?

[ncik-roberts]

Yes, happily changes to test files are picked up by -no-rebuild.